// ml systems · agentic ai · infrastructure
I build the rails and guardrails
for autonomous AI.From the model to the metal._
Reinforcement-learning systems, open-source agent tooling, and the infrastructure underneath.
autonomous agents on a leash they can’t loosen see how it works →
Selected work
01Agent Constitution & master-loop→A safety framework — and the controller that enforces it — for running fleets of autonomous coding agents.02interlock→A deterministic governance gate that decides which AI-agent pull requests can merge on their own.03tailored→An open-source engine that wraps a stochastic language model in deterministic gates, so every generated application pack is testable and repeatable.04Vault→A private vault where a sourcing autopilot grades roles twice a day and a human decides what to send, with the public tailored engine as the gate at its core.05grounded→An open-source hybrid retriever, vector search plus a knowledge graph, that cites every answer to what it found or abstains.06triagepilot→An in-app feedback loop where a language model filters, dedupes and diagnoses bug reports against the real source, and a schema gate decides what it is allowed to conclude.07delegate-skill-pack→The grill, plan, TDD-in-a-worktree, review and sweep workflow as a portable, agent-agnostic skill pack that any capable agent can pick up and run.08Seedstitch→A seeded generative engine that turns an image into a one-of-one embroidery design, and stitches the exact same design every time you give it the same seed.09Structural weight optimiser→A finite-element solver and a meta-heuristic optimiser, both hand-written in MATLAB, that search a steel frame toward minimum weight — every candidate validated against Abaqus.10Hokm RL agent→A reinforcement-learning agent that plays Hokm, a four-player hidden-hand card game, reasoning under partial observability.
About
A senior structural engineer — two master’s degrees, both with distinction — who kept reaching for code, until building software became the work. Today I build ML systems and agentic-AI tooling, and ship the whole thing end-to-end: model, interface, and the infrastructure it runs on.
the full story →Writing
Loops, all the way upWhat the levels of abstraction in an agent loop mean, how far up they reach, and the one loop none of them can close.Jun 2026Governing AI-agent pull requests: building interlockWhy I built a deterministic gate instead of another AI reviewer — and what "a fuse, not a judge" means in practice.Jun 2026
all writing →Contact
Building something interesting?
I’m always happy to talk shop — ML, agents, or the messy reality of shipping software.
[email protected]